Nous Research Says China's AI Model Political Bias Is Only a Thin Layer, Cut From 22% to 0.9%

At a glance

AI research company Nous Research released 'Arete-27B,' a model it trained further from Alibaba's open model Qwen3.8-27B. The share of answers that favor the Chinese government's position fell from 22% to 0.9%. The research team said this bias is shallow and sits in a relatively thin layer rather than deep in the model.

Key points

1. 1

   In the freedom of expression category, the share of biased answers dropped from 60% to 1%.

2. 2

   The evaluation used 803 questions across seven fields, and other AI models were scored on how much accurate fact they contained.

3. 3

   Training used only supervised fine-tuning (shaping the model by showing it example correct answers), without reinforcement learning.

4. 4

   The research team itself set the evaluation questions and scoring criteria and acknowledged that this creates limitations.

Why it matters

The same method could be used to change the answer tendencies of other AI models, so organizations using open models need to check not only performance but also how they were trained and where their data came from.

Also reported by

• 스타트업레시피中 AI 큐웬 정부 옹호 답변 줄인 아레테-27B 공개